What is ZeeSpec?
ZeeSpec is Vishal Mysore's method for spec-driven development: 60 questions in six groups (what, where, when, who, why, how), ten each, meant to be answered in about an hour so that an agent has no undefined behaviour left to guess. A small Node.js CLI scaffolds eight Markdown templates in a .spec/ folder, assembles them into AGENTS.md and copies the prompt to the clipboard.
The browser app at vishalmysore.github.io/zeespec/app has three tabs. Spec Builder walks the questions and can draft AGENTS.md with a local WebGPU model. EPIC Score grades a spec, a public GitHub repo or pasted text against a heuristic version of the EPIC rubric from an October 2026 arXiv paper: 40 practices across 10 dimensions, scored out of 30. SpecGuard re-creates SpecGuardBench in miniature: 16 JavaScript tasks testing whether a model's validation harness rejects a buggy implementation while accepting two correct ones. The repository has 14 commits and no releases; the last commit was on 16 November 2026.
Is ZeeSpec worth it?
The useful idea here is the SpecGuard tab, not the method. It shows, in a browser and without a GPU, the failure that a paper named SpecGuardBench documents: ask a model to write tests from a reference implementation and it mirrors the reference, bugs included. Of the three built-in baselines, mirror the reference scores 0% faithful, replaying the existing test scores 0%, and only a hand-written oracle reaches 100%. Every harness runs in a network-less worker with a time limit, and you can edit and re-run it. As a demonstration of why agent-written tests approve bugs, it is clearer than most prose on the subject.
The EPIC Score tab is more modest than its badge suggests. It is a heuristic re-implementation of the paper's rubric, not the authors' tool, and the app says so; the paper's finding that top-band specs spent 11.8% of commits on bug fixes against 20.4% for the bottom band is a correlation across 114 repositories.
The 60 questions are good questions, the kind a careful tech lead asks: what must never be stored, who must never act, when should the system block. The one-hour promise is softened by the project's own disclaimer.
The catch: the README's install command, `npx spec-cli init`, fetches an unrelated npm package called spec-cli, published by another author in 2024 and pointing at a different repository; ZeeSpec itself is not on npm. Running it executes a stranger's code. Clone the repository and run bin/spec.js instead. The project has 18 stars and 14 commits.
Our take: treat ZeeSpec as a worksheet and a classroom demo. GitHub's Spec Kit is the production tool for the same job.
Where it earns its place
An engineering manager running an internal workshop on agent-written tests
Show developers why tests generated from code miss that code's bugs
Open the SpecGuard tab, switch on mirror trap mode and run the three baselines live: the oracle scores 100% faithful, the other two 0%. Free, nothing to install, and every harness runs sandboxed in the browser.
A solo developer starting a small booking app with Claude Code
Write a spec complete enough that the agent stops inventing rules
Work through the 60 questions in Spec Builder, export AGENTS.md, then check it in the EPIC Score tab and paste its fix-it prompt into the agent. For the CLI, clone the repo rather than running npx.
Before you commit
- Setup
- Open the web app in a browser with WebGPU for the local model, or clone the repository and run `node bin/spec.js init`. Do not run `npx spec-cli`: that name on npm belongs to an unrelated package.
- Integrations
- Produces AGENTS.md for any agent that reads it; EPIC Score reads Spec Kit, Kiro, ZeeSpec and AGENTS.md layouts from public GitHub repos; WebLLM for the optional local model.
- Data and privacy
- The app states that nothing leaves your browser; the model download and GitHub repo fetches are the only network traffic. SpecGuard harnesses run in network-less workers. MIT licence.
- Limits
- SpecGuard has 16 JavaScript tasks against the paper's 291 Python tasks, so scores are not comparable; EPIC scoring is heuristic; no npm package or tagged releases.
What’s good
- SpecGuard demonstrates mirrored-test failures clearly
- Browser app runs locally, with network-less test workers
- 60 concrete questions expose undefined behaviour
- Honest about heuristic scoring and its limits
What’s not
- README's npx command installs an unrelated npm package
- EPIC score is a heuristic, not the paper's tool
- Tiny project: 18 stars, 14 commits, no releases
- Local model download of about 1 to 2 GB
Best for
Tech leads and teachers who want a structured spec checklist and a vivid demonstration of how agents write tests that pass bugs.
ZeeSpec alternatives
ZeeSpec: common questions
- Is ZeeSpec worth it?
- ZeeSpec's questions are a sound checklist and its SpecGuard tab is a sharp little teaching tool, but the CLI's install line points at someone else's package. We rate it 3.3 out of 5.
- How much does ZeeSpec cost?
- Free and open source (MIT). The web app runs in your browser; its optional local model downloads about 1 to 2 GB once.
- What are the downsides of ZeeSpec?
- README's npx command installs an unrelated npm package; ePIC score is a heuristic, not the paper's tool; tiny project: 18 stars, 14 commits, no releases; local model download of about 1 to 2 GB.
- Who is ZeeSpec best for?
- Tech leads and teachers who want a structured spec checklist and a vivid demonstration of how agents write tests that pass bugs.
- Is ZeeSpec safe for company data?
- The app states that nothing leaves your browser; the model download and GitHub repo fetches are the only network traffic. SpecGuard harnesses run in network-less workers. MIT licence.
- What are the best ZeeSpec alternatives?
- The closest alternatives we have reviewed are Spec Kit and Agent Skills (Addy Osmani).
Answers checked Nov 21, 2026.